基于文体计量和机器学习的英俄语仇恨言论自动识别研究OACHSSCD

[Purpose]This study aims to investigate the common linguistic characteristics and strategic differences of cross-lingual hate speech,providing empirical insights for the development of an interpretable multilingual threat intelligence monitoring system.[Method]Based on a corpus of approximately 2.2 million English and Russian online hate speech tokens,this study combines stylometric analysis to examine common features across five dimensions:words,sentences,readability,complexity,and sentiment.This study extracted latent stylistic structures through exploratory factor analysis,validated the distribution of features in the feature space using UMAP,and quanti-fied the contribution of each micro-indicator to classification decisions by combining machine learning with the SHAP attribution model,thereby establishing a core indicator system for identifying cross-lingual hate speech.[Result/Conclusion]A total of 52 features were i-dentified as common linguistic characteristics for distinguishing hate speech,including 15 lexical features,19 sentence features,8 readabil-ity features,7 complexity features,and 3 sentiment features.The experiments revealed significant differences in stylistic strategies between English and Russian:English tends toward structural complexity to increase the difficulty of censorship,while Russian tends toward sim-plicity and accessibility to accelerate online dissemination.Attribution analysis confirmed that the proportion of negative sentiment and proper nouns serve as the core decision-making basis for cross-lingual identification.This reveals the fundamental commonality of cross-lingual cyber threats in targeting specific audiences and inciting negative emotions.It provides interpretable core feature support for the au-tomatic identification of online public opinion risks and early warning of perceived threats,offering empirical data to enhance national cy-bersecurity governance capabilities.